AR | Build Notes

Baidu just released ERNIE 5.5 Flash Base: a 280B sparse MoE with only 14B active parameters, trained on 33.5T tokens. It's published as a pure base model, before any mid- or post-training. That gives researchers a clean checkpoint to run their own fine-tuning, RL and ablations on. #AI #LLM #OpenWeights

Thinking Machines just released Inkling 2 Small: a 594B-parameter MoE with only 17B active per token. It scores 31 on the Artificial Analysis Intelligence Index and is live on Tinker. Apache 2.0 weights are planned, which would make a 594B sparse model fully open. #AI #LLM #OpenSource

Together AI's new Together Compiler beat the best InferenceX configurations by up to 2.6x tokens per second, and it needed under 24 hours per model to do it. It's an agent-driven loop that keeps re-optimizing the inference forward pass as traffic patterns shift, so tuning isn't a one-time hand-built job anymore. Agents tuning the stack that serves the models. #AI #LLM

Ripple-65B Preview, a 65B dense open-weight model, scores 80.5 on Terminal-Bench 2.1 and edges out US open models up to 15x its size. Ricursive Intelligence shipped it under Apache 2.0, so the weights are free to use commercially. A dense 65B beating much larger models on agentic terminal work is a strong sign that parameter count isn't the whole story. #AI #OpenSource

Axion 1.0 is a 35B open agent model trained in just three stages: two light SFT passes, then RL. No mid-training, no DPO, no on-policy distillation. It's competitive with leading commercial models on several agentic benchmarks. The bigger deal: the whole recipe is public. Data construction, filtering, verification, prompts and training code, plus the training dataset. #AI #OpenSource #LLM

Anthropic released FlashProve: a zkSNARK prototype that proves a reply came from committed model weights, without revealing the weights. It proves Llama 4 Maverick at 110 tokens/second on a single H100, with 19 MB proofs. Apache 2.0. Verifiable inference, so you can check an API served the model it promised. #AI #zkML

Hydrogen-1.0 is a 753B open-weight model for chemistry, materials science and life sciences, and it ships under the MIT License. Deep Principle built it by post-training Z.ai's GLM-5.3 on domain data, lab records and agent discovery trajectories. They report a FrontierScience-Research win over GPT-6-Astra. Open weights for scientific discovery, at frontier scale. #AI #OpenSource

Zyphra's ZHA retrofits a frozen LLM with tree-indexed sparse attention using just 4B tokens of distillation, and it matches dense attention on reasoning and 256K-context retrieval benchmarks. No pretraining from scratch. The hierarchical tree index picks which tokens to attend to, a step toward true subquadratic attention for long context. #AI #LLM

tandem.cpp pushes Qwen3.8-Flash-Next to 97 tok/s on a single RTX 4090 + 64 GB DDR4. That's 3.3× stock llama.cpp (29.8 tok/s) on a 12400F box. It's a llama.cpp patch set that pins the hottest MoE experts in VRAM and has the CPU handle the rest. Works with any MoE llama.cpp supports: gpt-oss-120b gets 1.85×, and others land around 1.4-2.1×. Prompt processing on 64k contexts is 1.4-1.8× faster too. #LocalLLM #llamacpp

Nace AI open-sources Drex Max v1, a 27B multimodal decision model that scores 67.52 and runs locally on Apple Silicon. A 27B model on a Mac means no API bill, no data leaving your machine, and a decision engine you can actually own. Weights are out, so the benchmark claim is yours to check. #AI #OpenSource #LocalLLM